Introducing AIOZ-GDANCE: A New Dataset for Group Dance Generation

Dancing is an essential part of human culture and communication. With the growth of social media, creating and sharing dance videos has become extremely popular. As a result, millions of dance videos are viewed daily on various online platforms. Researchers have been interested in finding ways to create natural dance movements based on music. While there has been progress in this area, most studies have focused on solo dancing. The task of creating dance movements for groups of dancers is still a significant challenge.

#The Challenge of Group Dance Generation

Creating dance motions for a group is more complex than for a single dancer. In group dancing, there may be different choreographies for each dancer while still keeping in sync with the music. Moreover, dancers often interact with each other physically, and ensuring that their movements match without interfering with one another adds to the difficulty. Most available datasets focus on solo dance videos and do not capture crucial elements of group dance, like synchronization and interaction among dancers.

For effective research in group dance generation, a large dataset is crucial. Many existing datasets are small or rely on expensive motion capture technology, making it hard to gather a diverse range of data. Some efforts have used algorithms to extract dance movements from videos available online, but these datasets were designed primarily for solo dancers.

#Presenting AIOZ-GDANCE

To address these shortcomings, we introduce a new dataset called AIOZ-GDANCE. This dataset is created for group dance generation and includes videos of groups dancing, along with their music. Our dataset features a more extensive collection than most others, supporting a variety of dance styles and music genres.

A unique aspect of our dataset is its semi-automatic labeling process. This involves people helping to label the data, ensuring high quality. AIOZ-GDANCE consists of hours of videos that showcase group dancing in real-world situations, with paired music and 3D dance movements.

#The Method Behind the Dataset

To develop a large-scale dataset, we decided against using traditional motion capture systems due to their cost and complexity. Instead, we collected group dance videos from public platforms like YouTube and Facebook. All videos were processed to ensure standard resolution and frame rates.

To track the movements of all dancers in the videos, we employed advanced tracking technology to get the bounding boxes of each dancer. This step is vital for reconstructing the movements accurately. We also used Pose Estimation methods to generate the initial 2D poses of each dancer. While some errors appear in this phase due to issues like motion blur, we manually corrected these inaccuracies.

By applying a method to fit 3D motion to each dancer's movements, we ensured that the representation of each dancer was captured accurately. This involved optimizing all dancers' motions simultaneously to create a coherent group motion, taking into account their physical interactions.

#How the Dataset Can Be Used

With AIOZ-GDANCE, we hope to inspire future research in group dance generation. While single dancer choreography has received considerable attention, group dancing still requires more exploration.

In addition to group dance generation, our dataset can help with various tasks such as human pose tracking and motion analysis. It can also be beneficial in areas like dance education and behavior analysis. Researchers are encouraged to explore these and other potential applications of the dataset.

#Audio-Driven Group Dance Generation

Our goal is to generate group motion sequences based on a given music audio input and initial positions of dancers. The objective is for the generated dance movements to be in harmony with the music, which involves maintaining rhythm and style. The dancers' movements should also be coherent, ensuring they do not collide or interfere with each other.

#The Architecture of Group Dance Generator

Our system uses a special architecture that takes the music sequence and initial dancer positions as input. The music is processed to extract meaningful features, which are then encoded to capture the essential parts of the audio.

We generate initial poses for dancers by combining music features with their starting positions. A key part of our Motion Generation involves ensuring that all dancers' movements are synchronized and consistent with the music. As part of this, we use attention mechanisms to consider how dancers relate to each other in space and time.

#Experimentation and Results

We built our model to train on the AIOZ-GDANCE dataset. By doing so, we could measure how well our method performed in creating coherent group dance movements. We compared our approach with existing methods, which typically focused on solo dancing.

Our experiments revealed that our method, especially when utilizing the attention mechanism, outperformed alternatives. It was better at handling issues like dance intersections, where dancers might collide or overlap during their movements.

We also looked at how the number of dancers affected the performance. Some metrics showed consistency across different numbers of dancers, indicating that our method could generate reliable dance movements regardless of how many dancers were involved. However, we observed that generating more dancers also increased the likelihood of intersection issues.

Next, we examined how different dance styles presented unique challenges in group dance generation. Styles such as Zumba and Aerobics were easier for our model to replicate due to their straightforward movements. In contrast, styles like Bollywood and Samba, which involve more intricate movements, posed greater difficulties in capturing accurately.

#Conclusion

In summary, we have presented AIOZ-GDANCE, a large dataset designed for group dance generation based on music. This dataset aims to support research in this area, providing the necessary resources for developing new methods for creating dance movements for groups. Our work shows promise for advancing studies related to audio-driven group choreography and offers a pathway for others to explore further applications.